When the USPTO is faced with a patent application of uncertain validity, which way should it err? How do the costs of erroneous patent grants compare with the costs of erroneous patent denials? Steve Yelderman provides an insightful new take on these questions in The Value of Accuracy in the Patent System (forthcoming in the University of Chicago Law Review). In short, he concludes that erroneous patent grants are more harmful than previously assumed, but only when "the unsatisfied patentability requirement is one that seeks to influence a mutually exclusive choice"—i.e., whether it incentivizes invention over non-invention, disclosure over non-disclosure, licensing over reinvention, etc.
Under the conventional account, if one assumes that the legal criteria of patentability are set correctly (an important caveat!), then social welfare is maximized by awarding patents to patentable inventions (true positives) and denying patents to unpatentable inventions (true negatives). Erroneous denials (false negatives) create a cost of decreased ex ante incentives for other inventors (IFN), while erroneous grants (false positives) and correct grants create ex post costs like deadweight loss (C). If the probability that a given application is patentable is q, then it should only be granted at the threshold q > C/IFN. (Note that it doesn't matter whether IFN is included as a cost for false negatives or a benefit for true positives; you get to the same result.) If the ex post costs of patents are low relative to their incentive benefits, they should be granted even for inventions that are unlikely to meet patentability standards, and as the relative weight of these factors reverses, patents should only be awarded when it seems quite likely that they are deserved.
Yelderman's key insight is that this analysis ignores another important cost of erroneous grants: false positives also reduce ex ante incentives because they often narrow the expected difference between inventing and not inventing. This framework can be illustrated as follows, with IFP shown in blue: